I am hired to make systems smaller.

Retail · vending hardware

The demo had been slipping for weeks. Fourteen days later the first machine was working on the street.

Retail through vending hardware under state traceability rules for precious metals: every transaction has to be provable, not merely recorded. The system was eight months old, eight people had built it, and rewriting it was not an option.

The pain was not the quality of the work: nobody could take the product from the first step to the last and say it was ready. Effort was going into what would matter in a year, while the first release was underestimated.

One question against everything: what stops us going live tomorrow? First we wrote down what the system already did — you cannot safely remove what nobody has written down. Then about 80% of what existed was removed and the rest finished. What had been one inseparable lump came apart into thirteen parts that could be worked on one at a time.

Joined on 17 June, first test run on 30 June, and from 1 July the first machine was live, restocked by a contractor’s courier. That is the time to launch, not the time of the work: the bulk of the system landed over the two months that followed, with the business already running on it.

• 8 months old at entry
• team of 8
• ~80% of code removed
• 13 separate parts
• 14 days to the first machine live

Details in conversation

Banking · regulatory reporting and internal systems

Five people cannot hold fifty different worlds in their heads.

A team of about five people is responsible for roughly fifty applications, eight to ten in flight at once. The thing to count was never the applications but the switches between them: each system had its own habits, its own way of being built and released, its own shape.

The variety looked like a technical problem and was a human one. Much of the estate was built between 2005 and 2010, and every system still carries the habits of the year it was born in. But that was not the pain. The pain was that every entry into a colleague’s system cost a day of remembering — and five people stop coping long before fifty applications run out.

What gets standardized is the pain, not the technology: one way to build a system, one way to release it, one way to set it up. Whatever did not hurt enough was left alone — parts of the estate still talk to each other the old way, and leaving them is cheaper than bringing them into line.

How you can tell the pain is gone: when the environment changed what it demanded, the applications themselves barely moved. What changed was one thin layer, and it changed the same way in every application it touched. The first change was made by the most expensive person on the team; the rest were repeated by the cheapest. Work that would once have meant rebuilding fifty applications cost almost nothing. The rest follows from the same place: more than 60% of the estate rewritten, moved or built the new way, and incidents have stayed quiet.

Details in conversation